Abstract
A template-driven framework supports development and deployment of large ecosystems of machine learning models. A representative subset of models is selected from a model fleet by clustering model descriptors across multiple ecosystem dimensions. In a template iteration phase, candidate techniques are evaluated on the representative subset and a multi-model optimization process identifies template-level hyperparameters that satisfy an aggregate performance criterion while constraining regressions. Qualified techniques and hyperparameters are encoded into a versioned modular standard model template having standardized interfaces for architecture and, in some embodiments, feature, data, and training/serving components. In a model iteration phase, many individual production models are instantiated from the template using model-specific inputs such as objectives, pipelines, features, and serving constraints, without per-model architectural redesign. The approach enables scalable technique propagation and consistent model construction across diverse products and hardware targets.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Standard Model Template", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10642